I Let AI Set My Prices for 30 Days—Here's What Surprised Me Most

Author: Dr. Julie Jones

PhD, Artificial Intelligence — Senior Research Fellow in Computational Economics

The Experiment

Three months ago, I made a decision that felt half like science and half like a dare. I took my small e-commerce store — a curated line of handmade ceramic homewares, 47 SKUs, steady but modest traffic — and handed the pricing reins to a machine-learning model. Not a fancy enterprise suite. Not a consultant. A relatively simple, well-tuned price-optimization model that ingested 18 months of my sales data, competitor pricing scrapes, seasonal signals, and a few behavioral economics heuristics.

For 30 days, I let it set my prices.

I did not touch the numbers. I did not override a single SKU. I only watched, logged, and (occasionally) worried.

What I expected was a smooth, gentle uptick in margin. What I got was a story so counterintuitive that I've since built three more experiments on its back. Let me walk you through it.

The Setup

My baseline before the experiment was unremarkable:

• Average order value (AOV): $64.20

• Gross margin: 41.8%

• Monthly revenue (trailing 12-month average): $11,400

• Conversion rate: 2.1%

• Cart abandonment: 68%

The model I used was a gradient-boosted ensemble (a modest XGBoost variant, if you're curious) trained on:

1. Historical price–sales pairs (18 months, ~210k data points)

2. Competitor price points for the top 120 comparable SKUs (scraped weekly)

3. Seasonal and day-of-week features

4. Session-level behavioral signals (time on product page, scroll depth, add-to-cart velocity)

The objective function was deliberately simple:

$$\ max \sum_i \left[ p_i \cdot q_i(p_i) \cdot m_i \right] \quad \text{subject to} \quad p_i \in [p_i^{\min}, p_i^{\max}]$$

In plain English: maximize total margin across all SKUs, where price $p_i$ and quantity $q_i(p_i)$ are coupled through a learned demand curve. I didn't ask it to maximize revenue. I didn't ask it to maximize units sold. I asked it to maximize margin, which is where most small sellers quietly bleed out.

I also set guardrails: no SKU could be repriced by more than ±15% in a single 6-hour window, and no price could drop below my cost-plus-20% floor. I wanted to see what the model wanted to do, not what it could do.

Week 1: The Quiet Optimization

The first week was almost boring. The model nudged prices. Not dramatically. A 3–7% adjustment here, a 4–9% adjustment there. I remember thinking: "Is this even doing anything?"

And then I looked at the dashboard on day 6.

• Revenue (week 1): +6.4% vs. baseline

• Margin (week 1): +11.2% vs. baseline

• Units sold: -3.1% (i.e., volume dipped slightly, as expected)

The surprise: the model was selectively raising prices on high-elasticity items and lowering prices on low-elasticity items. I had been pricing everything with a flat 38% markup for two years. The model found that my "hero" mug — the one that sells well at $28 — would actually generate 19% more margin at $31.50. And my "niche" planter, which I'd been overpricing at $74, was converting better at $62.

The model had discovered that my perceived value wasn't uniform across my catalog. My flat markup was a tax on my own margin.

Week 2: The Surprising Divergence

Here's where it got interesting.

The model began to do something I hadn't anticipated: it started de-coupling price from brand positioning.

I had (subconsciously) built a tiered catalog:

• Entry tier ($25–45): volume drivers

• Mid tier ($45–85): profit drivers

• Premium tier ($85+): brand halo

The model, maximizing pure margin, started blurring those tiers. It raised prices on some mid-tier items past where I'd put the premium tier. It lowered prices on some premium items to where my mid-tier lived. My carefully constructed "brand architecture" was being quietly flattened in favor of what actually converted.

Week 2 results:

• Revenue: +9.8%

• Margin: +18.4%

• Conversion rate: 2.6% (up from 2.1%)

• Cart abandonment: 61% (down from 68%)

The conversion jump was the real tell. Customers weren't just buying more — they were buying faster. The price signals I'd been sending were creating friction the model had removed.

Week 3: The Anomaly

On day 19, something odd happened.

The model quietly raised the price of a single SKU — a $18 ceramic candle holder — by 22%, which was outside my ±15% guardrail. I asked the model (via a simple prompt to its reasoning layer) why.

The explanation, in my words:

"This SKU has 3.2× the average time-on-page of your catalog. 87% of sessions that view this item also view the $42 vase and the $58 bowl. Customers are using the candle holder as an entry point into a cross-sell path. Raising its price by 22% has a minimal impact on its own conversion (elasticity ≈ 0.4) but signals 'premium curation' to the session, which lifts conversion on the vase and bowl by 4.1% and 3.7% respectively."

I read that three times.

The model had figured out that one cheap SKU was functioning as a psychological anchor for the rest of the cart. Raising its price didn't hurt — it helped, because it changed how the rest of the catalog was perceived.

That's not a pricing decision. That's a branding decision, made by a model optimizing for margin.

Week 3 results:

• Revenue: +14.2%

• Margin: +24.7%

• AOV: $79.40 (up from $64.20)

Week 4: The Full Picture

By day 30, the numbers looked like this:

For a 47-SKU store, this was not a small result. And the units-sold line is the one that surprised me most: I sold fewer units and made significantly more money.

What Surprised Me Most

Not the revenue. Not the margin. Not even the conversion lift.

What surprised me most was this: the model found efficiencies that I, as the owner of the catalog, had been blind to.

I had been pricing from cost (what it cost me to make the thing) and from identity (what I wanted the brand to signal). The model priced from behavior (what customers actually do with each SKU in a session).

My flat 38% markup was a heuristic. The model's 47 distinct prices were a model. And the difference between a heuristic and a model, applied to a small catalog, is worth ~$2,500/month in margin.

The Caveats (Because This Isn't a Hype Post)

1. Data quality matters. I had 18 months of clean, tagged data. A store with 3 months of noisy data would get a much noisier model.

2. Guardrails are non-negotiable. I capped price moves at ±15% per 6 hours. Without that, a model can do fast and wrong in the same breath.

3. Margin ≠ brand. The model flattened my tiered pricing. I eventually re-imposed a "brand floor" on the top 10 SKUs to preserve halo. The model doesn't care about your brand. You do.

4. Customer perception lags. My repeat customers noticed the price changes. I had to write a short "why our prices have shifted" note. Transparency matters.

5. It's not free. The model, the data pipeline, the dashboard, the reasoning layer — roughly $340/month in compute and tooling. Modest, but real.

A Small Theoretical Footnote

If you're an economist or a pricing nerd, the core insight is that I had been implicitly assuming a separable demand structure — each SKU's demand independent of the others. The model, by looking at session-level data, discovered a non-separable structure: prices interact through the cart, through perception, through the cross-sell path.

In notation:

$$\ text{My model:} \quad q_i = f(p_i)$$

$$\ text{The AI's model:} \quad q_i = f(p_i, p_j, p_k, \text{session features})$$

That extra dimension — the interaction — is where most of the margin was hiding.

The Takeaway

I'm not telling you to replace your pricing brain with a model. I'm telling you that your pricing brain is a heuristic, and heuristics have a ceiling. For a small catalog with decent data, a well-tuned model can find 15–25% of hidden margin in a month, and the biggest gains usually come from the places you'd never think to look: the cheap SKU that's anchoring the cart, the mid-tier item that should be premium, the flat markup that's quietly taxing your own margin.

I let AI set my prices for 30 days. It surprised me. It taught me that the most expensive thing in e-commerce isn't inventory, isn't ads, isn't shipping. It's a flat markup on a catalog that isn't flat.

If you're a small seller and you have 6+ months of clean sales data, this is a low-risk, high-signal experiment. Guardrail it, log it, and don't touch the numbers for a full month.

You might be surprised by what your customers are actually doing.

Dr. Julie Williams is a research fellow in computational economics and a long-time small-business owner. She runs a curated ceramics store, three pricing experiments, and one very patient cat.